The AI Wrapper Business in 2026: Why Most Fail and What the Winners Did Differently

What is an AI wrapper? A working definition for 2026

An AI wrapper is a product whose core value depends on calls to a model someone else built. The founder adds a prompt, an interface and a billing page on top of an API from OpenAI, Anthropic or Google, and sells the result. The AI wrapper business was the default startup move of 2023 and 2024 because it shipped fast and cost almost nothing to start.

By 2026 the word covers two very different companies. A thin wrapper is a feature that is one model upgrade away from irrelevance. A durable one uses the model as a component inside a workflow, a data asset or a distribution engine. Cursor and Perplexity both began on models they did not own, and critics have used the wrapper label on both. The label describes where a company started, not where it can end up.

That distinction is the whole story for US founders and operators weighing AI business ideas right now. The question is no longer whether you can build on a model API. It is what still belongs to you the day a model vendor ships your feature for free.

Why most AI wrapper startups fail

Estimates of how many AI startups fail run from 80% to more than 90%, and nearly all trace back to secondary sources. A figure repeated across blogs attributes a roughly 80% wrapper failure rate by the end of 2026 to CB Insights and Gartner, but AETW could not locate a primary report that states it. Treat the exact percentage as a directional signal. The mechanisms behind it are easier to verify, and there are three.

The first is platform absorption, and Jasper is the cleanest case study. Its recurring revenue doubled to about $80M in 2022 on plans that mostly cost $80 a month, and it raised at a $1.5B valuation. Then ChatGPT launched in November 2022. Sacra's analysis found customers swapping $50-a-month subscriptions for free and $20-a-month ChatGPT plans. Jasper cut its 2023 revenue projections by as much as 30%, laid off staff in July 2023 and replaced its CEO.

The second is the margin squeeze. A wrapper pays its supplier on every request. ICONIQ's January 2026 State of AI report, as summarized by Beancount, found inference averaging 23% of revenue at scaling-stage AI B2B companies, with 84% of them reporting gross margin erosion of six points or more tied to AI infrastructure. Bessemer puts typical AI gross margins at 50% to 60%, against 80% to 90% for traditional SaaS. Heavy users make it worse. One operator guide estimates they can cost 50 to 100 times more to serve than light users on the same plan, which turns flat-rate pricing into a bet against your own best customers.

The third is the absence of switching costs. If the product is a prompt and an interface, a competitor can copy it quickly and a model vendor can absorb it in a release. Cluely, covered next, had a free open source lookalike circulating around its biggest launch.

Cluely: the loudest AI wrapper business, and its limits

Cluely sells a desktop assistant that watches a user's screen and listens to calls, then feeds answers in real time for interviews, sales calls and meetings. It launched in 2025 with a deliberately provocative pitch built around the motto "Cheat on everything," raised a $5.3M seed from Abstract Ventures and Susa Ventures, then a $15M Series A led by Andreessen Horowitz. For a company that young, that is a serious run, and it came from distribution, not technology.

The numbers are where it gets complicated. In July 2025, CEO Roy Lee told TechCrunch that ARR had doubled in a week to about $7M after an enterprise launch, up from a claimed $3M. In March 2026 he admitted on X that the $7M figure was not real. One summary of that admission puts actual ARR near $5.2M, split between roughly $2.7M consumer and $2.5M enterprise. That is still real revenue for a startup this young. The lesson is about trust: in a market where ARR is self-reported and unaudited, a fabricated figure costs credibility that distribution cannot buy back.

The moat question was visible from the start. TechCrunch noted in July 2025 that the real-time notetaker could be easy to replicate, and that a startup called Pickle had released Glass, a free open source product similar to Cluely. Attention also decays. By DataForSEO's US numbers, searches for the brand peaked near 110,000 a month in September 2025 and ran around 40,500 in August 2026.

Cluely is best read as proof that distribution is a real asset in the AI wrapper business, and a fragile one. Rage-bait marketing buys awareness. It does not buy retention, switching costs or proprietary data.

What the winning AI wrapper companies did differently

The survivors did not avoid building on other people's models. They made sure the model was the least valuable part of the product. Four cases show four different routes.

Cursor moved down the stack. It reportedly crossed $100M in ARR in January 2025, $1B in November 2025 and about $2B in February 2026, then roughly $4B by June 2026, figures that come from press reports rather than audited financials. Early on it rented third-party models and ran at negative gross margins, per CRV's write-up. It then shipped its own Composer model in late 2025 and routed work between in-house and cheaper external models to reach slightly positive gross margin. SpaceX agreed to acquire Anysphere, Cursor's parent, for $60B in stock on June 16, 2026, and the deal reportedly closed on August 14.

Harvey went vertical. The legal AI product sold to law firms and in-house teams passed $100M in ARR in August 2025 and reportedly sits between $300M and $400M by September 2026, depending on the source. Sacra reported that Harvey scrapped its fine-tuned legal model once frontier reasoning models commoditized legal reasoning. That is vertical AI working as intended: when the model becomes a commodity, value moves to workflow, integrations, compliance and sales into a regulated buyer. Its valuation reportedly rose from $11B in March 2026 to roughly $15.5B to $16B in September.

Lovable went after speed and habit. The Swedish vibe-coding startup grew from $200M in ARR in November 2025 to about $600M by September 2026 and raised $400M at a $13.3B valuation on August 12, 2026. The company said net dollar retention exceeded 100% in February 2026, a company-reported figure. The caveat is mix: enterprise revenue was only about $20M when ARR hit $400M, so most of the base is still consumers and small teams, the same segment that drained out of Jasper.

Perplexity bet on owning the interface and the habit. ARR passed $450M in March 2026 per the Financial Times, up from about $100M a year earlier, and the company was last priced at $20B in September 2025. It also faces copyright suits from News Corp and Encyclopaedia Britannica, a reminder that building on other people's content creates its own risks.

The pattern across all four is the same. Each owns something the model vendor cannot ship in a release: its own model layer, a regulated workflow, a distribution habit, or some mix. That is what people mean by an AI moat, and it is why the same API call can sit under a $60B company and a dead startup.

The 2026 picture: bigger winners and a thinner middle

The AI wrapper business in 2026 looks like a barbell. At one end, a handful of companies are growing at speeds with no precedent in software: Cursor roughly doubled ARR in four months, and Lovable tripled it in about ten months. At the other end, generic tools that launched into the 2023 gold rush face falling prices, tighter margins and customers who now treat AI writing, summarizing and chat as things that come free.

Margins are the dividing line. The average AI product gross margin sits near 52%, up from 41% in 2024, per ICONIQ data cited by Beancount. The strongest operators are working it up by routing tasks to cheaper models, moving to usage-based or credit pricing, or building their own models, as Cursor did.

Valuations tell the same story from the investor side. Cursor's $60B price on roughly $4B of ARR is about 15x, down from 30x or more when critics measured the same number against $2B. Harvey trades near 39x to 52x depending on which ARR figure you use. Perplexity's $20B price on $450M works out to about 44x. Those are growth-stock multiples for companies with moats, not for products without them.

Treat all of it with discipline. Nearly every ARR number here is company-reported or press-sourced and unaudited, and Cluely already showed how far a headline figure can drift from reality. The vertical bet is not risk-free either. One analysis of Harvey flags that incumbents such as Thomson Reuters or LexisNexis could ship competing AI at scale.

What to watch next: whether Cursor keeps its independence inside SpaceX, whether Lovable can build a real enterprise business beyond its roughly $20M of enterprise revenue, and whether the next price cut from a model vendor takes another layer of thin products with it.

A stress test before you build one

• Run the ship-it-tomorrow test: if OpenAI, Google and Anthropic all shipped your exact feature natively, would your users stay? If not, you have a feature, not a company.
• Price against cost per customer, not blended gross margin. Heavy users can cost far more than light ones, so meter usage or sell credits.
• Name the asset that compounds: proprietary data, workflow integration, a distribution habit, or your own model layer. If you cannot name one, the AI moat is missing.
• Track 90-day retention and net dollar retention from day one. Jasper and Cluely both show that consumer attention decays faster than it builds.
• Report ARR with explicit definitions. Cluely's retraction shows that credibility with investors, press and customers is itself an asset.
• Pick a niche where general models are structurally weak, such as regulated work that demands audit trails. That is where Harvey built its lead.

Sources